xref: /freebsd/sys/netpfil/ipfw/ip_dn_private.h (revision a0409676120c1e558d0ade943019934e0f15118d)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2010 Luigi Rizzo, Riccardo Panicucci, Universita` di Pisa
5  * All rights reserved
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 /*
30  * internal dummynet APIs.
31  *
32  * $FreeBSD$
33  */
34 
35 #ifndef _IP_DN_PRIVATE_H
36 #define _IP_DN_PRIVATE_H
37 
38 /* debugging support
39  * use ND() to remove debugging, D() to print a line,
40  * DX(level, ...) to print above a certain level
41  * If you redefine D() you are expected to redefine all.
42  */
43 #ifndef D
44 #define ND(fmt, ...) do {} while (0)
45 #define D1(fmt, ...) do {} while (0)
46 #define D(fmt, ...) printf("%-10s " fmt "\n",      \
47         __FUNCTION__, ## __VA_ARGS__)
48 #define DX(lev, fmt, ...) do {              \
49         if (dn_cfg.debug > lev) D(fmt, ## __VA_ARGS__); } while (0)
50 #endif
51 
52 MALLOC_DECLARE(M_DUMMYNET);
53 
54 #ifndef __linux__
55 #define div64(a, b)  ((int64_t)(a) / (int64_t)(b))
56 #endif
57 
58 #define DN_LOCK_INIT() do {				\
59 	mtx_init(&dn_cfg.uh_mtx, "dn_uh", NULL, MTX_DEF);	\
60 	mtx_init(&dn_cfg.bh_mtx, "dn_bh", NULL, MTX_DEF);	\
61 	} while (0)
62 #define DN_LOCK_DESTROY() do {				\
63 	mtx_destroy(&dn_cfg.uh_mtx);			\
64 	mtx_destroy(&dn_cfg.bh_mtx);			\
65 	} while (0)
66 #if 0 /* not used yet */
67 #define DN_UH_RLOCK()		mtx_lock(&dn_cfg.uh_mtx)
68 #define DN_UH_RUNLOCK()		mtx_unlock(&dn_cfg.uh_mtx)
69 #define DN_UH_WLOCK()		mtx_lock(&dn_cfg.uh_mtx)
70 #define DN_UH_WUNLOCK()		mtx_unlock(&dn_cfg.uh_mtx)
71 #define DN_UH_LOCK_ASSERT()	mtx_assert(&dn_cfg.uh_mtx, MA_OWNED)
72 #endif
73 
74 #define DN_BH_RLOCK()		mtx_lock(&dn_cfg.uh_mtx)
75 #define DN_BH_RUNLOCK()		mtx_unlock(&dn_cfg.uh_mtx)
76 #define DN_BH_WLOCK()		mtx_lock(&dn_cfg.uh_mtx)
77 #define DN_BH_WUNLOCK()		mtx_unlock(&dn_cfg.uh_mtx)
78 #define DN_BH_LOCK_ASSERT()	mtx_assert(&dn_cfg.uh_mtx, MA_OWNED)
79 
80 SLIST_HEAD(dn_schk_head, dn_schk);
81 SLIST_HEAD(dn_sch_inst_head, dn_sch_inst);
82 SLIST_HEAD(dn_fsk_head, dn_fsk);
83 SLIST_HEAD(dn_queue_head, dn_queue);
84 SLIST_HEAD(dn_alg_head, dn_alg);
85 
86 #ifdef NEW_AQM
87 SLIST_HEAD(dn_aqm_head, dn_aqm); /* for new AQMs */
88 #endif
89 
90 struct mq {	/* a basic queue of packets*/
91         struct mbuf *head, *tail;
92 	int count;
93 };
94 
95 static inline void
96 set_oid(struct dn_id *o, int type, int len)
97 {
98         o->type = type;
99         o->len = len;
100         o->subtype = 0;
101 }
102 
103 /*
104  * configuration and global data for a dummynet instance
105  *
106  * When a configuration is modified from userland, 'id' is incremented
107  * so we can use the value to check for stale pointers.
108  */
109 struct dn_parms {
110 	uint32_t	id;		/* configuration version */
111 
112 	/* defaults (sysctl-accessible) */
113 	int	red_lookup_depth;
114 	int	red_avg_pkt_size;
115 	int	red_max_pkt_size;
116 	int	hash_size;
117 	int	max_hash_size;
118 	long	byte_limit;		/* max queue sizes */
119 	long	slot_limit;
120 
121 	int	io_fast;
122 	int	debug;
123 
124 	/* timekeeping */
125 	struct timeval prev_t;		/* last time dummynet_tick ran */
126 	struct dn_heap	evheap;		/* scheduled events */
127 
128 	/* counters of objects -- used for reporting space */
129 	int	schk_count;
130 	int	si_count;
131 	int	fsk_count;
132 	int	queue_count;
133 
134 	/* ticks and other stuff */
135 	uint64_t	curr_time;
136 	/* flowsets and schedulers are in hash tables, with 'hash_size'
137 	 * buckets. fshash is looked up at every packet arrival
138 	 * so better be generous if we expect many entries.
139 	 */
140 	struct dn_ht	*fshash;
141 	struct dn_ht	*schedhash;
142 	/* list of flowsets without a scheduler -- use sch_chain */
143 	struct dn_fsk_head	fsu;	/* list of unlinked flowsets */
144 	struct dn_alg_head	schedlist;	/* list of algorithms */
145 #ifdef NEW_AQM
146 	struct dn_aqm_head	aqmlist;	/* list of AQMs */
147 #endif
148 
149 	/* Store the fs/sch to scan when draining. The value is the
150 	 * bucket number of the hash table. Expire can be disabled
151 	 * with net.inet.ip.dummynet.expire=0, or it happens every
152 	 * expire ticks.
153 	 **/
154 	int drain_fs;
155 	int drain_sch;
156 	uint32_t expire;
157 	uint32_t expire_cycle;	/* tick count */
158 
159 	int init_done;
160 
161 	/* if the upper half is busy doing something long,
162 	 * can set the busy flag and we will enqueue packets in
163 	 * a queue for later processing.
164 	 */
165 	int	busy;
166 	struct	mq	pending;
167 
168 #ifdef _KERNEL
169 	/*
170 	 * This file is normally used in the kernel, unless we do
171 	 * some userland tests, in which case we do not need a mtx.
172 	 * uh_mtx arbitrates between system calls and also
173 	 * protects fshash, schedhash and fsunlinked.
174 	 * These structures are readonly for the lower half.
175 	 * bh_mtx protects all other structures which may be
176 	 * modified upon packet arrivals
177 	 */
178 #if defined( __linux__ ) || defined( _WIN32 )
179 	spinlock_t uh_mtx;
180 	spinlock_t bh_mtx;
181 #else
182 	struct mtx uh_mtx;
183 	struct mtx bh_mtx;
184 #endif
185 
186 #endif /* _KERNEL */
187 };
188 
189 /*
190  * Delay line, contains all packets on output from a link.
191  * Every scheduler instance has one.
192  */
193 struct delay_line {
194 	struct dn_id oid;
195 	struct dn_sch_inst *si;
196 	struct mq mq;
197 };
198 
199 /*
200  * The kernel side of a flowset. It is linked in a hash table
201  * of flowsets, and in a list of children of their parent scheduler.
202  * qht is either the queue or (if HAVE_MASK) a hash table queues.
203  * Note that the mask to use is the (flow_mask|sched_mask), which
204  * changes as we attach/detach schedulers. So we store it here.
205  *
206  * XXX If we want to add scheduler-specific parameters, we need to
207  * put them in external storage because the scheduler may not be
208  * available when the fsk is created.
209  */
210 struct dn_fsk { /* kernel side of a flowset */
211 	struct dn_fs fs;
212 	SLIST_ENTRY(dn_fsk) fsk_next;	/* hash chain for fshash */
213 
214 	struct ipfw_flow_id fsk_mask;
215 
216 	/* qht is a hash table of queues, or just a single queue
217 	 * a bit in fs.flags tells us which one
218 	 */
219 	struct dn_ht	*qht;
220 	struct dn_schk *sched;		/* Sched we are linked to */
221 	SLIST_ENTRY(dn_fsk) sch_chain;	/* list of fsk attached to sched */
222 
223 	/* bucket index used by drain routine to drain queues for this
224 	 * flowset
225 	 */
226 	int drain_bucket;
227 	/* Parameter realted to RED / GRED */
228 	/* original values are in dn_fs*/
229 	int w_q ;		/* queue weight (scaled) */
230 	int max_th ;		/* maximum threshold for queue (scaled) */
231 	int min_th ;		/* minimum threshold for queue (scaled) */
232 	int max_p ;		/* maximum value for p_b (scaled) */
233 
234 	u_int c_1 ;		/* max_p/(max_th-min_th) (scaled) */
235 	u_int c_2 ;		/* max_p*min_th/(max_th-min_th) (scaled) */
236 	u_int c_3 ;		/* for GRED, (1-max_p)/max_th (scaled) */
237 	u_int c_4 ;		/* for GRED, 1 - 2*max_p (scaled) */
238 	u_int * w_q_lookup ;	/* lookup table for computing (1-w_q)^t */
239 	u_int lookup_depth ;	/* depth of lookup table */
240 	int lookup_step ;	/* granularity inside the lookup table */
241 	int lookup_weight ;	/* equal to (1-w_q)^t / (1-w_q)^(t+1) */
242 	int avg_pkt_size ;	/* medium packet size */
243 	int max_pkt_size ;	/* max packet size */
244 #ifdef NEW_AQM
245 	struct dn_aqm *aqmfp;	/* Pointer to AQM functions */
246 	void *aqmcfg;	/* configuration parameters for AQM */
247 #endif
248 };
249 
250 /*
251  * A queue is created as a child of a flowset unless it belongs to
252  * a !MULTIQUEUE scheduler. It is normally in a hash table in the
253  * flowset. fs always points to the parent flowset.
254  * si normally points to the sch_inst, unless the flowset has been
255  * detached from the scheduler -- in this case si == NULL and we
256  * should not enqueue.
257  */
258 struct dn_queue {
259 	struct dn_flow ni;	/* oid, flow_id, stats */
260 	struct mq mq;	/* packets queue */
261 	struct dn_sch_inst *_si;	/* owner scheduler instance */
262 	SLIST_ENTRY(dn_queue) q_next; /* hash chain list for qht */
263 	struct dn_fsk *fs;		/* parent flowset. */
264 
265 	/* RED parameters */
266 	int avg;		/* average queue length est. (scaled) */
267 	int count;		/* arrivals since last RED drop */
268 	int random;		/* random value (scaled) */
269 	uint64_t q_time;	/* start of queue idle time */
270 #ifdef NEW_AQM
271 	void *aqm_status;	/* per-queue status variables*/
272 #endif
273 
274 };
275 
276 /*
277  * The kernel side of a scheduler. Contains the userland config,
278  * a link, pointer to extra config arguments from command line,
279  * kernel flags, and a pointer to the scheduler methods.
280  * It is stored in a hash table, and holds a list of all
281  * flowsets and scheduler instances.
282  * XXX sch must be at the beginning, see schk_hash().
283  */
284 struct dn_schk {
285 	struct dn_sch sch;
286 	struct dn_alg *fp;	/* Pointer to scheduler functions */
287 	struct dn_link link;	/* The link, embedded */
288 	struct dn_profile *profile; /* delay profile, if any */
289 	struct dn_id *cfg;	/* extra config arguments */
290 
291 	SLIST_ENTRY(dn_schk) schk_next;  /* hash chain for schedhash */
292 
293 	struct dn_fsk_head fsk_list;  /* all fsk linked to me */
294 	struct dn_fsk *fs;	/* Flowset for !MULTIQUEUE */
295 
296 	/* bucket index used by the drain routine to drain the scheduler
297 	 * instance for this flowset.
298 	 */
299 	int drain_bucket;
300 
301 	/* Hash table of all instances (through sch.sched_mask)
302 	 * or single instance if no mask. Always valid.
303 	 */
304 	struct dn_ht	*siht;
305 };
306 
307 /*
308  * Scheduler instance.
309  * Contains variables and all queues relative to a this instance.
310  * This struct is created a runtime.
311  */
312 struct dn_sch_inst {
313 	struct dn_flow	ni;	/* oid, flowid and stats */
314 	SLIST_ENTRY(dn_sch_inst) si_next; /* hash chain for siht */
315 	struct delay_line dline;
316 	struct dn_schk *sched;	/* the template */
317 	int		kflags;	/* DN_ACTIVE */
318 
319 	int64_t	credit;		/* bits I can transmit (more or less). */
320 	uint64_t sched_time;	/* time link was scheduled in ready_heap */
321 	uint64_t idle_time;	/* start of scheduler instance idle time */
322 
323 	/* q_count is the number of queues that this instance is using.
324 	 * The counter is incremented or decremented when
325 	 * a reference from the queue is created or deleted.
326 	 * It is used to make sure that a scheduler instance can be safely
327 	 * deleted by the drain routine. See notes below.
328 	 */
329 	int q_count;
330 
331 };
332 
333 /*
334  * NOTE about object drain.
335  * The system will automatically (XXX check when) drain queues and
336  * scheduler instances when they are idle.
337  * A queue is idle when it has no packets; an instance is idle when
338  * it is not in the evheap heap, and the corresponding delay line is empty.
339  * A queue can be safely deleted when it is idle because of the scheduler
340  * function xxx_free_queue() will remove any references to it.
341  * An instance can be only deleted when no queues reference it. To be sure
342  * of that, a counter (q_count) stores the number of queues that are pointing
343  * to the instance.
344  *
345  * XXX
346  * Order of scan:
347  * - take all flowset in a bucket for the flowset hash table
348  * - take all queues in a bucket for the flowset
349  * - increment the queue bucket
350  * - scan next flowset bucket
351  * Nothing is done if a bucket contains no entries.
352  *
353  * The same schema is used for sceduler instances
354  */
355 
356 /* kernel-side flags. Linux has DN_DELETE in fcntl.h
357  */
358 enum {
359 	/* 1 and 2 are reserved for the SCAN flags */
360 	DN_DESTROY	= 0x0004, /* destroy */
361 	DN_DELETE_FS	= 0x0008, /* destroy flowset */
362 	DN_DETACH	= 0x0010,
363 	DN_ACTIVE	= 0x0020, /* object is in evheap */
364 	DN_F_DLINE	= 0x0040, /* object is a delay line */
365 	DN_DEL_SAFE	= 0x0080, /* delete a queue only if no longer needed
366 				   * by scheduler */
367 	DN_QHT_IS_Q	= 0x0100, /* in flowset, qht is a single queue */
368 };
369 
370 /*
371  * Packets processed by dummynet have an mbuf tag associated with
372  * them that carries their dummynet state.
373  * Outside dummynet, only the 'rule' field is relevant, and it must
374  * be at the beginning of the structure.
375  */
376 struct dn_pkt_tag {
377 	struct ipfw_rule_ref rule;	/* matching rule	*/
378 
379 	/* second part, dummynet specific */
380 	int dn_dir;		/* action when packet comes out.*/
381 				/* see ip_fw_private.h		*/
382 	uint64_t output_time;	/* when the pkt is due for delivery*/
383 	struct ifnet *ifp;	/* interface, for ip_output	*/
384 	struct _ip6dn_args ip6opt;	/* XXX ipv6 options	*/
385 	uint16_t iphdr_off;	/* IP header offset for mtodo()	*/
386 };
387 
388 /*
389  * Possible values for dn_dir. XXXGL: this needs to be reviewed
390  * and converted to same values ip_fw_args.flags use.
391  */
392 enum {
393 	DIR_OUT =	0,
394 	DIR_IN =	1,
395 	DIR_FWD =	2,
396 	DIR_DROP =	3,
397 	PROTO_LAYER2 =	0x4, /* set for layer 2 */
398 	PROTO_IPV4 =	0x08,
399 	PROTO_IPV6 =	0x10,
400 	PROTO_IFB =	0x0c, /* layer2 + ifbridge */
401 };
402 
403 extern struct dn_parms dn_cfg;
404 //VNET_DECLARE(struct dn_parms, _base_dn_cfg);
405 //#define dn_cfg	VNET(_base_dn_cfg)
406 
407 int dummynet_io(struct mbuf **, struct ip_fw_args *);
408 void dummynet_task(void *context, int pending);
409 void dn_reschedule(void);
410 struct dn_pkt_tag * dn_tag_get(struct mbuf *m);
411 
412 struct dn_queue *ipdn_q_find(struct dn_fsk *, struct dn_sch_inst *,
413         struct ipfw_flow_id *);
414 struct dn_sch_inst *ipdn_si_find(struct dn_schk *, struct ipfw_flow_id *);
415 
416 /*
417  * copy_range is a template for requests for ranges of pipes/queues/scheds.
418  * The number of ranges is variable and can be derived by o.len.
419  * As a default, we use a small number of entries so that the struct
420  * fits easily on the stack and is sufficient for most common requests.
421  */
422 #define DEFAULT_RANGES	5
423 struct copy_range {
424         struct dn_id o;
425         uint32_t	r[ 2 * DEFAULT_RANGES ];
426 };
427 
428 struct copy_args {
429 	char **start;
430 	char *end;
431 	int flags;
432 	int type;
433 	struct copy_range *extra;	/* extra filtering */
434 };
435 
436 struct sockopt;
437 int ip_dummynet_compat(struct sockopt *sopt);
438 int dummynet_get(struct sockopt *sopt, void **compat);
439 int dn_c_copy_q (void *_ni, void *arg);
440 int dn_c_copy_pipe(struct dn_schk *s, struct copy_args *a, int nq);
441 int dn_c_copy_fs(struct dn_fsk *f, struct copy_args *a, int nq);
442 int dn_compat_copy_queue(struct copy_args *a, void *_o);
443 int dn_compat_copy_pipe(struct copy_args *a, void *_o);
444 int copy_data_helper_compat(void *_o, void *_arg);
445 int dn_compat_calc_size(void);
446 int do_config(void *p, int l);
447 
448 /* function to drain idle object */
449 void dn_drain_scheduler(void);
450 void dn_drain_queue(void);
451 
452 #ifdef NEW_AQM
453 int ecn_mark(struct mbuf* m);
454 
455 /* moved from ip_dn_io.c to here to be available for AQMs modules*/
456 static inline void
457 mq_append(struct mq *q, struct mbuf *m)
458 {
459 #ifdef USERSPACE
460 	// buffers from netmap need to be copied
461 	// XXX note that the routine is not expected to fail
462 	ND("append %p to %p", m, q);
463 	if (m->m_flags & M_STACK) {
464 		struct mbuf *m_new;
465 		void *p;
466 		int l, ofs;
467 
468 		ofs = m->m_data - m->__m_extbuf;
469 		// XXX allocate
470 		MGETHDR(m_new, M_NOWAIT, MT_DATA);
471 		ND("*** WARNING, volatile buf %p ext %p %d dofs %d m_new %p",
472 			m, m->__m_extbuf, m->__m_extlen, ofs, m_new);
473 		p = m_new->__m_extbuf;	/* new pointer */
474 		l = m_new->__m_extlen;	/* new len */
475 		if (l <= m->__m_extlen) {
476 			panic("extlen too large");
477 		}
478 
479 		*m_new = *m;	// copy
480 		m_new->m_flags &= ~M_STACK;
481 		m_new->__m_extbuf = p; // point to new buffer
482 		_pkt_copy(m->__m_extbuf, p, m->__m_extlen);
483 		m_new->m_data = p + ofs;
484 		m = m_new;
485 	}
486 #endif /* USERSPACE */
487 	if (q->head == NULL)
488 		q->head = m;
489 	else
490 		q->tail->m_nextpkt = m;
491 	q->count++;
492 	q->tail = m;
493 	m->m_nextpkt = NULL;
494 }
495 #endif /* NEW_AQM */
496 
497 #endif /* _IP_DN_PRIVATE_H */
498